docs: lock inference review — prompts, Call 1/2, feat tags

Restore CONTENT-WOO, add PROMPTS.md, and align PRD/architecture/API
with inference decisions: prompt files on disk, 0.7 confidence gate,
species constrained to Woo leaves, Call 2 prose-only + templates,
feat-* tags, and inference_enabled gating.

Slab capture app PRD.txt

@@ -20,16 +20,20 @@ publish pipeline.
 1. No calibration photos — user manually enters SKU, length, thickness.
 2. First photo is the measurement photo (must be top-down, lens parallel to broad side).
 3. All photos go through background removal + 80% frame-fill crop.
-4. Output: transparent PNGs, minimum 1600 px on the smaller dimension, aspect ratio preserved.
+4. Output: transparent PNGs, minimum 1600 px on the smaller dimension,
+   cropped to 3:4 portrait (W:H 3:4 = 4:3 portrait) to match the WooCommerce
+   template; configurable in settings.
 5. Green or black sheet — algorithm auto-detects background color from image borders.
 6. WooCommerce categories synced from store: Species (1 required, leaf-only), Wood Category (1+ required, leaf-only).
 7. WooCommerce attributes: Edge Type (1 required), Figure (1+ required), Grade (1+ required), Thickness (1 required, rounded up to nearest Woo attribute range).
-8. WooCommerce tags: fig-* only (1+ required, e.g., fig-chat = Chatoyant).
+8. WooCommerce tags: fig-* and feat-* only (1+ required, e.g., fig-chat, feat-makcr).
+   WooCommerce is the authoritative source; the design does not duplicate tag lists.
 9. Two-call inference: Call 1 (vision, wood expert) → Call 2 (text, content gen). Both configurable.
-10. Call 1 runs automatically after calibration. User reviews/corrects before triggering Call 2.
+10. Call 1 runs automatically after calibration **only if** inference is enabled.
+    User reviews/corrects before triggering Call 2.
 11. Inference timeout >45s or non-200 → error with retry. User can fill fields manually.
 12. Woo product status set to "published" by default (configurable in settings).
-13. All inference prompts stored server-side, editable in settings with reset-to-default.
+13. Inference prompts are server-side text files, read fresh on each invocation.
 14. Brand voice + GEO context injected into Call 2 prompt.
 15. WooCommerce credentials (username + App Password) stored client-side in settings.
 16. Inference endpoint: configurable OpenAI-compatible API (user provides base URL, key, model).
@@ -94,8 +98,8 @@ publish pipeline.
 - **FR3:** App detects background color from image borders (green or black sheet).
 - **FR4:** First photo: BG removal - find longest dimension - compute scale
   (longest_px / user-length) - compute area (sqft) - bdft = sqft * thickness / 12.
-- **FR5:** All photos: BG removal - crop to 80% frame fill - save as transparent PNG,
-  minimum 1600 px on smaller dimension, aspect ratio preserved.
+- **FR5:** All photos: BG removal - crop to 80% frame fill - save as transparent
+  PNG, minimum 1600 px on smaller dimension, 3:4 portrait crop (configurable).
 - **FR6:** User can retake the first photo or abort and start fresh.
 
 ### Dimensions and calculations
@@ -124,22 +128,38 @@ publish pipeline.
 ### Inference (two-call architecture)
 - **FR22:** Call 1 (vision): auto-run after calibration. Input: all inventory photos
   (downscaled to 1024px max, using original pre-crop photos), user metadata, synced
-  Woo taxonomy. Output: species, wood categories, edge type, figure, grade,
-  character/inclusions/voids/checks with severity, top 3 wood categories with
-  confidence. If no species match above threshold, omit and require manual selection.
-- **FR23:** Call 2 (text): user-triggered after Call 1 curation. Input: deterministic
-  data + Call 1 results + brand voice + GEO context. Output: title, description,
-  short description.
-- **FR24:** Inference prompts stored server-side, editable in settings, reset-to-default available.
+  Woo taxonomy. Output: species, wood categories, edge type, figure, grade —
+  selected from Woo taxonomy, each with a confidence score. The model is
+  constrained to choose from the synced Woo taxonomy lists.
+- **FR22a:** Confidence threshold: 0.7. If the model's confidence for any field
+  is below 0.7, the field is left blank and requires manual user selection.
+  Threshold is configurable in settings for tuning.
+- **FR22b:** Species bias: the prompt includes the full Woo species leaf list,
+  constraining the model to select from known species only. This prevents
+  hallucinated species and significantly improves accuracy.
+- **FR22c:** Free-text observations: character, inclusions, voids, and checks
+  are free-text observations. The server maps them to Woo `feat-*` product
+  tags after the call returns. WooCommerce is the authoritative source for
+  feat-* tag definitions; the design does not duplicate them.
+- **FR23:** Call 2 (text): user-triggered. Can use curated Call 1 results OR
+  manually-entered taxonomy. The LLM writes only descriptive prose
+  (character, color, grain, use cases). The server assembles the final title,
+  short description, and full description from deterministic templates with the
+  LLM prose injected. Dimensions are never generated by the LLM.
+- **FR24:** Inference prompts are server-side text files, read fresh on each
+  invocation. A reset-to-default script restores the shipped defaults. Prompts
+  are not user-facing settings; they are tuned by the owner directly.
 - **FR25:** Progress indicator during inference calls (latency expected).
-- **FR26:** Timeout >45s or non-200 - visible error with retry option.
+- **FR26:** Timeout >45s or non-200 - visible error with retry option. Retry is
+  available if the inference endpoint was previously tested as connected.
+  Intermittent failures are expected in low-connectivity environments.
 - **FR27:** All inference results editable; no automatic publish of LLM output.
 - **FR28:** Inference endpoint configured via settings: base URL, API key, model name.
   Test connection verifies text + vision support.
 
 ### Photo normalization and output
 - **FR29:** Background removal: chroma-key for green/black, fallback to U2Net matting.
-- **FR30:** Crop to 80% frame fill, slab centered, aspect ratio preserved,
+- **FR30:** Crop to 80% frame fill, slab centered, 3:4 portrait aspect,
   minimum 1600 px on smaller dimension.
 - **FR31:** Output as transparent PNG (ideal for WooCommerce).
 - **FR32:** Store both original and processed images. Purge after successful publish.